Pooling Data Across Microarray Experiments Using Different Versions of Affymetrix Oligonucleotide Arrays

1. Pooling Data Across Microarray Experiments Using Different Versions of Affymetrix Oligonucleotide Arrays Jeffrey S. Morris Li Zhang, Chunlei Wu, Keith Baggerly and Kevin Coombes UT MD Anderson Cancer Center Houston, TX, USA

2. Combining Information across Microarray Studies • Many publicly available microarray data sets • Can combine information across studies to: • Validateresults from individual studies • Find intersection of differentially expressed genes • Build model using one study, validate using another • Discover new biological insightsby analyses pooling data across studies. • Potential for increased statistical power • Important since many individual studies are underpowered.

3. Pooling Data across Studies • Challenge: In general, microarray data from different studies not comparable • Clinical differences • Different study populations • Technical differences • Laboratory differences: sample collection and storage, microarray protocol • Different platforms: cDNA/oligo, different versions of same technology (e.g. Affy chips)

4. Pooling Data across Studies • Approaches in Existing Literature: • Include study effects in model • Gene-specific study effects • SVD, Distance-weighted Discrimination Drawback: First-order corrections not enough • Model unitless summary measures • standardized log fold-change • t-statistics • probabilities of +/0/- expression Drawback: Implicit assumptions about comparability of clinical populations across studies

5. Pooling Data across Studies • Sequence-related reasons for incomparability of raw expression levels across platforms: • Cross-hybridization • RNA degradation (near 5’ end) • Probe validity – map to RefSeq? • Alternative splicing • It may be possible that, by taking these into account, we can obtain more comparable raw expression levels to use in pooled analyses • Our focus: combining information across different versions of Affymetrix genechips

6. Overview of Affymetrix GeneChips • Probes:25-base sequences from gene of interest • Probesets: set of probes corresponding to same gene. • Obtained from current sequence information in GenBank, Unigene, RefSeq • Generations of human chips: • HuGeneFL: 5600 genes, 20 probes/gene • U95Av2: 10,000 genes, 16 probes/gene • U133A: 14,500 genes, 11 probes/gene

7. Example: CAMDA Lung Cancer Data • CAMDA: “Critical Assessment of Microarray Data Analysis”: annual conference at Duke University • CAMDA 2003: Two studiesrelating gene expression data to survival in lung cancer patients • Harvard(Bhattacharjee, et al. 2001) • 124 lung adenocarcinoma samples • Michigan (Beer, et al. 2002) • 86 lung adenocarcinoma samples • GOAL: Pool data across studies to identify prognostic genes for lung cancer.

8. Pooling Information across Studies Harvard patients – worse prognosis

9. Pooling Information across Chip Types • Michigan: HuGeneFl Chip • 6,633 probe sets -- 20 probe pairs each • Harvard: HG_U95Av2 Chip • 12,453 probe sets – 16 probe pairs each • Problems: • Different genes • Incomparable expression levels

10. “Partial Probeset” Method “Partial Probesets” 1. Identify “matching probes” 2. Recombine into new probesets based on UNIGENE clusters, which we refer to as “partial probesets” 3. Eliminate any probesets containing just one or two probes • Note: Any quantification method can subsequently be used (MAS, dChip, RMA, PDNN) HuGeneFL : HG_U95Av2: … … Matching Probes

11. Pooling Information across Chip Types

12. Quantification of Expression Levels • Gene expressions quantified by applying Li’s PDNN model to our partial probesets • Uses probe sequence info to predict patterns of specific and nonspecific hybridization intensities • Allows borrowing of strength across probe sets • Model is not overparameterized – O(N probesets) • See Zhang, et al. (2003) Nature Biotech for further details on method and comparison

13. Detecting Outliers • Log-scale plots to detect outliers • Large spot detected on 4 Michigan chips L54 L88 L89 L90 • Other outliers: 6 from Michigan, 2 Harvard • Other preprocessing (remove low expr./normalize) • Matching clinical/microarray data for 200 patients (124 H, 76 M)

14. Assessing Our Method for Combining Information Across Chip Types • “Partial Probeset” method appears to give comparable expression levels across chip types.

15. Assessing our Method for Combining Information across Chip Types • Median “partial probeset” size is 7, vs. 16 or 20 Loss of precision? • No evidence of significant precision loss

16. Assessing “Partial Probeset” Method • Agreement in relative quantifications across samples

17. Assessing “Partial Probeset” Method • Agreement in relative quantifications across samples • Less variable genes worse

18. Assessing “Partial Probeset” Method • Agreement in relative quantifications across samples • Less variable genes worse • Eliminate genes with sd<0.20 or r<0.90

19. Assessing “Partial Probeset” Method • Agreement in relative quantifications across samples • Less variable genes worse • Eliminate genes with sd<0.20 or r<0.90 • 1,036 genes

20. Identifying Prognostic Genes • Preprocess raw microarray data • Outlier Detection, Normalization, Quantification, Remove Some Genes? • Left with “n-by-p” matrix of expression levels for p genes on n microarrays. • Identify which genes are correlated to outcome of interest • Perform standard statistical test for each gene – obtain (permutation) p-values • Find “cutpoint” on p-values to declare significance that accounts for multiplicities.

21. Identifying Prognostic Genes • After preprocessing: • 1036 genes, 200 samples • Identify genes related to survival • After adjusting for known clinical predictors • Provide prognostic information on survival above and beyond clinical predictors

22. Identifying Prognostic Genes: Cox Regression Modeling • Hazard : (t) ~ Prob(X<t +t | X>t ) • Cox Model: i(t) = 0(t) exp(Xi  ) • Xi= Vector of covariates for subject i •  = Vector of regression coefficients • Key Assumption: Proportional Hazards • Hazard ratio between subjects with different covariates does not vary over time. • i(t )/k(t ) = exp{ (Xi-Xk) } • Exp() =Change in hazard per unit change in X

23. Identifying Prognostic Genes: Cox Regression Modeling • Best Clinical Model:

24. Identifying Prognostic Genes • Series of 1036 multivariable Cox models fit to identify prognostic genes. Each model contained: • Study (Michigan=-1, Harvard=1). • Age (continuous factor). • Stage (early=0/late=1). • Probeset (log intensity value as continuous factor). • Exact p-values for each probeset computed using permutation approach • By using multivariate modeling, we search for genes offering prognostic information beyondclinical predictors

25. Identifying Prognostic Genes: BUM Method • No prognostic genes  pvals Uniform • Prognostic genes  smaller pvals • Fit Beta-Uniform mixture to histogram of p-values – “BUM” method • (Pounds and Morris, 2003 Bioinformatics) • Method can be used to identify prognostic genes while controlling FDR

26. Results • Histogram suggests there are some significant probesets • FDR=0.20 corresponds pval cutoff of 0.0024 (BUM, Pounds and Morris 2003) • 26 probesets flagged as significant

27. Selected Flagged Genes

28. Selected Flagged Genes

29. Selected Flagged Genes

30. Selected Flagged Genes

31. Selected Flagged Genes

32. Selected Flagged Genes

33. Selected Flagged Genes

34. Selected Flagged Genes

35. Selected Flagged Genes

36. Selected Flagged Genes

37. Results • Our gene list has almost no overlap with other publications of these data. Reasons: • We addressed a different research question • Us: ID Genes offering prognostic info beyond clinical • Michigan: Univariate Cox models fit; results used to construct dichotomous “risk index” • Harvard: Cluster analysis done; clusters linked to survival; found genes driving the clustering • Pooling across studies yieldedsignificant gains in statistical power. • Most genes (17/26) in our study are not flagged if we analyze 2 data sets separately (i.e. no pooling)

38. CAMDA 2003 Results We Won!

39. Limitations of Partial Probeset Method • Worked well for combining across HuGeneFL/ U95Av2 • ~25% probes from HuGeneFL on U95Av2, with 4,101 probesets • Not enough matching probes for use with U95Av2/U133A • ~6% of probes from U95Av2 also on U133A, with only 628 probesets • Requiring matching probes strong criterion, maybe weaker criterion would suffice?

40. Alternative Splicing Diagram of C2GnT I gene organization and different mRNA variants of this gene that are differentially expressed across tissue types. From Falkenberg, et al. (2003) Glycobiology 13(6), 411-418.

41. Full-Length Transcript Based Probesets • New probeset definition (FLTBP): probes match the same set of full-length mRNA sequences • Procedure • Construct comprehensive library of full-length mRNA transcript sequences from RefSeq and HinvDB • For each probe, identify all matching full-length transcripts using Blast program U95Av2: 15% matched no sequence, 33% matched multiple seq. U133A: 18% matched no sequence, 38% matched multiple seq. • Group probes with same matched target lists (FLTBPs) U95Av2: 23,972 probesets, U133A: 14,148 probesets

42. Full-Length Transcript Based Probesets • Matching across chip types: • 9,642 FLTBPs match across U95Av2 and U133A • Affymetrix has their own method for mapping their probesets across arrays – 9,480 pairs of probesets (only about ½ map the same way as FLTBPs) • Example: Lung cancer cell line data • 28 cell lines, each hybridized onto both U95Av2 and U133A arrays. • Paired design suggests any differences between paired measurements due to technical, not biological, sources. • Different quantification methods (PDNN, RMA, MAS, dChip)

43. Results • Density estimate of chip-to-chip correlations for each gene • Positive shift for FLTBP suggests better correlations • Improvement greatest for PDNN • Correlation still not perfect

44. Example: Sample Gene 1 • Plot of probe signals for two chip types (Red=FLTBP) • Scatterplot of log-expression values for each sample across the two chip types (Black=all probes, Red=FLTBP) • Correlation across chips significantly improved with FLTBP

45. Example: Sample Gene 2 • Again, significantly higher correlation using FLTBP than using Affymetrix’ definition

46. PDNN RMA MAS5 dChip Results • Boxplot of chip-to-chip correlations (over genes) for each sample • PDNN resulted in higher correlations

47. Conclusions • New method for pooling info across studies using different versions of Affymetrix chips. • Recombine matched probes into new probesets using Unigene clusters. • Method appears to obtain comparable expression levels across chips without sacrificing much precision or significantly altering the relative ordering of the samples. • Worked well combining information across HuGeneFL/U95Av2, but not U95Av2/U133A

48. Conclusions • Discussed new probeset definition based on full-length transcript sequences. • Removes effect of known alternative splicing • Yields stronger between-chip correlations than Affymetrix standard definitions • Pooling information across studies is difficult – there is still more work to be done – but worth the effort.

49. References • Morris JS,Yin G, Baggerly KA, Wu C, and Zhang L (2005). Pooling Information Across Different Studies and Oligonucleotide Microarray Chip Types to Identify Prognostic Genes for Lung Cancer. Methods of Microarray Data Analysis IV, eds. JS Shoemaker and SM Lin, pp. 51-66,New York: Springer-Verlag. • Wu C, Morris JS, Baggerly KA, Coombes KR, Minna JD, and Zhang L (2005). A probe-to-transcripts mapping method for cross-platform comparisons of microarray data taking into account the effects of alternative splicing. Under review. • Morris JS,Wu C, Coombes KR, Baggerly KA, Wang J, and Zhang L (2006). Alternative Probeset Definitions for Combining Microarray Data Across Studies Using Different Versions of Affymetrix Oligonucleotide Arrays. To appear in Meta-Analysis in Genetics, edited by Rudy Guerra and David Allison, Chapman-Hall.